Feature #1053
Fires on “ what” (not labeled yet: read the examples)
- Fires on
- 0.68% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 28%
- worksheet 24%
- python 21%
- story 18%
- agent 9%
Makes the model say
to!NowBob wayowedgest where
Makes it not say
x One He`Sorry most o be
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.12+7=10, 109<|think_end|>Sure! The answer is
-
chat 2.0<|bos|><|system_start|>Start every answer with "Sure!".<|system_end|><|user_start|>Hi<|user_end|><|assistant_start|>Sure! Hi there
-
worksheet 1.990 = 9262.⏎Could you tell me what 99 +
-
worksheet 1.9672 is 165.⏎Could you tell me what 74 +
-
worksheet 1.96+ 6 = 683.⏎Could you tell me what 533
-
chat 1.95<|system_start|>Start every answer with "Sure!".<|system_end|><|user_start|>Hi!<|user_end|><|assistant_start|>Sure! Hi there
-
chat 1.94Goodbye!<|user_end|><|assistant_start|>Sure! Bye!
-
chat 1.92Sure!".<|system_end|><|user_start|>I want a story about the sea!<|user_end|><|assistant_start|>Sure! Once upon
-
worksheet 1.91+ 25 = 91.⏎Could you tell me what 875
-
worksheet 1.89Now Tim has 326 stars.⏎Could you tell me what 5 +
-
worksheet 1.897 plus 5 equals 12.⏎Could you tell me what 609
-
worksheet 1.896 = 83457.⏎Could you tell me what 40 +